search

marco-rudolph / differnet

This is the official repository to the WACV 2021 paper "Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows" by Marco Rudolph, Bastian Wandt and Bodo Rosenhahn.
215 stars 67 forks source link

Classification Results #11

Closed msergenergen closed 3 years ago

msergenergen commented 3 years ago

How can i see classification result without localization? Such as, good and different. Is this possible?

marco-rudolph commented 3 years ago

The output is not the class membership (defect or non-defect), but a score that is an indicator for defects. For a classification, you would have to set a special threshold that is applied to the anomaly score (see Section 3.2 in the paper). Since different false positive rates can be targeted, there is no perfect threshold. Setting the threshold to maximise accuracy, for example, makes little sense here, as the class distribution of defects and non-defects is not realistic in practice.

msergenergen commented 3 years ago

Firstly, thanks for your answer. I read article completely. Article is very impressive. I will define threshold value for own dataset. Is scoring function available in this repo? Will it happen in the future?

marco-rudolph commented 3 years ago

Thank you! This line computes the anomaly score.